Two-dimensional object detection system based on artificial intelligence inference

Abstract This study introduces an automated two-dimensional (2D) material detection system designed to identify and localize exfoliated thin 2D flakes using optical microscope images. The detection approach leverages the optical contrast produced by interference effects between the 2D material and its thin-film substrate—a technique traditionally relied upon for manual inspection. The system decomposes optical images acquired by a camera into individual RGB channels, quantifies the contrast ratio between the flake region and the surrounding substrate, and initially pinpoints candidate areas. To further refine detection, statistical features such as mean contrast difference, intensity uniformity, and local gradient behavior are analyzed, allowing the exclusion of multilayer fragments, tape residues, and other contaminants. However, relying solely on RGB contrast often results in misclassifications, especially when impurities share similar optical properties with true 2D flakes. To address this limitation, we comparatively investigated four CNN-based approaches, ResNet18, VGG11, MobileNetV2, and an image preprocessing pipeline termed Feat3 in this work, and identified ResNet18 as the most appropriate model in terms of overall classification performance. The CNN model learns the morphological signatures distinguishing authentic flakes from non-target artifacts. Through this novel approach, the system achieved a CNN patch-level classification accuracy > 97% on an independent test dataset, significantly improving robustness and reliability compared to contrast-based methods alone.